The goal of our research is to make AI systems more efficient, reliable, and secure across all software and hardware layers. Our current projects span agentic AI operations, intelligent model serving, hardware introspection, and hardware security. Our group is working on three main projects: LLexus, Switchcraft, and Project Sapphire (opens in new tab). We publish in top academic conferences including ISCA, SOSP, SIGCOMM, NSDI, and USENIX Security. Our work has been productized across several Microsoft groups (MS Foundry, Azure PhyNet, Office on the Web, Azure Powercapping, Azure EngOps), has shipped in Microsoft hardware (Azure Cobalt SoC), or has been adopted by industry (JEDEC).

Together, all our work reflects our group’s vision:

co-designing novel and intelligent software systems and hardware-aware techniques to improve the efficiency, reliability, scalability, and trustworthiness of next-generation AI systems.